Sample self-selection using dual teacher networks for pathological image classification with noisy labels.

Area of Science:

• Computer Science
• Artificial Intelligence
• Medical Imaging

Background:

• Deep neural networks (DNNs) require extensive high-quality labeled data for optimal performance.
• Noisy labels in datasets significantly degrade DNN accuracy, a critical issue in medical image analysis where labeled data is scarce.
• Existing methods to handle noisy labels often lead to underfitting.

Purpose of the Study:

• To develop a robust method for training DNNs with noisy labels, particularly in the medical domain.
• To mitigate the performance degradation caused by inaccurate labels in pathological image datasets.
• To improve the reliability and accuracy of DNN models in real-world applications with imperfect data.

Main Methods:

• A data sample self-selection strategy utilizing early stopping to identify and filter out noisy data samples.
• Implementation of knowledge distillation with dual teacher networks to stabilize the learning process of the student network.
• Leveraging model predictions from training history to curate cleaner datasets for retraining.

Main Results:

• The proposed method outperforms existing state-of-the-art techniques for handling noisy labels on both synthetic and real-world datasets.
• Achieved a notable increase in classification accuracy of 2.39% on the Chaoyang pathological image dataset.
• Demonstrated significant mitigation of the impact of noisy labels on model performance through data cleaning and retraining.

Conclusions:

• The novel approach effectively addresses the challenge of noisy labels in DNN training by combining early stopping and knowledge distillation.
• The method enhances model robustness and accuracy, offering a practical solution for medical image analysis.
• This strategy provides a pathway to more reliable AI models in domains with limited or imperfect data.